MULTIPATH-ENABLED PRIVATE AUDIO WITH NOISE
Anadi Chaman
Yu-Jeh Liu
Jonah Casebeer
Ivan Dokmanić
University of Illinois, Urbana Champaign
ICASSP 2019
[Paper]
[Code]
[Poster]

Fig. 1: Problem setup

In this work, we present a private audio communication system between a set of centrally coordinated loudspeakers and microphones in a reverberant room. The goal is to send audio messages from the speakers to the microphones while making sure that an eavesdropper at any other location does not understand these messages. We call the microphone locations 'focusing spots'.

This can be achieved if the signals emitted from speakers are designed such that the resulting sounds jam all locations other than the focusing spots. In this work, we have proposed two approaches that use noise to address this task. Unlike the conventional multi-zone soundfield reproduction systems whose performance is degraded in reverberant environments, our solutions work well precisely by exploiting echoes. Our experiments show that given the room impulse responses (RIRs) between the speakers and focusing spots, it is possible to achieve private audio communication with merely 6 speakers.

Our Solution

  1. Multichannel convolutional synthesis (MCCS) by noise: In this method, audio messages are communicated to intended focusing spots by emitting appropriately filtered white Gaussian noise signals from loudspeakers. The filters are constructed such that after echoing through specific sets of paths and time delays, these filtered random signals sum up coherently as they arrive at the target focusing points. This random 'scrambling' of filters ensures that only if they arrive at the intended locations, they correctly get descrambled
    Convolution of filters with noise signals can be interpreted as a scrambling process. Given a set of randomly drawn noise signals, the filters are constructed such that after convolution (scrambling)

  2. Noise in nullspace approach: This approach involves sending random noise from loudspeakers in addition to message signals, such that the noise signals add up to zero only at the intended locations while continuing to mask the messages everywhere else. This results in the interception of clean audio messages at the focusing spots while having low intelligibility at other locations. This technique is inspired from standard approaches to eavesdropper jamming in wireless communication. However, unlike prior works in wireless, this jamming scheme has been adapted to work with long convolutional channels that are common in room-acoustics.

Results

Consider a scenario where the system intends to send two audio signals to listeners Alice and Bob in a room, meanwhile ensuring that a third listener Eve can not eavesdrop on these signals. The locations of Alice, Bob and Eve in the room are randomly chosen. We perform numerical simulations and real-world experiments with this setting.


  1. Numerical simulation results
    2.
    Below are the two signals to be communicated.

Audio signal desired to reach Alice
Audio signal desired to reach Bob

We use our two proposed approaches to send these messages privately to Alice and Bob using 6 speakers. The Room impulse responses between the speakers and the two listeners are computed using Pyroomacoustics package. The following are the resultant audio signals heard by Alice, Bob and Eve.

Signals received MCCS approach Nullspace approach
Audio signal heard by Alice
Audio signal heard by Bob
Audio signal heard by Eve


  1. Real-world experiments
    2.
This experiment was performed in an office space of size 10 m x 6m using two Genelec 8030B and four Genelec 8010A loudspeakers. Microphones were placed at the locations of Alice, Bob and Eve. The room impulse responses used in our approach were measured using the exponential sine sweep technique.
This experiment was repeated twice each with a different speaker-user configuration and different audio signals.

First configuration

The following signals are to be communicated.
Audio signal desired to reach Alice
Audio signal desired to reach Bob



Signals received MCCS approach Nullspace approach
Audio signal heard by Alice
Audio signal heard by Bob
Audio signal heard by Eve



Second configuration

The following signals need to be communicated.
Audio signal desired to reach Alice
Audio signal desired to reach Bob



Signals received MCCS approach Nullspace approach
Audio signal heard by Alice
Audio signal heard by Bob
Audio signal heard by Eve

Paper and Citation

[Paper]  [ArXiv]

Citation
 
 Anadi Chaman, Yu-Jeh Liu, Jonah Casebeer, Ivan Dokmanić. Multipath-enabled Private Audio with Noise
 In IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) (pp. 685-689). 2019.